A Winner Takes All mechanism for automatic object extraction from multi-source data

Abstract

Automatic object extraction from multi-source aerial data is a desirable property for many activities, such as detecting 3D city model changes or updating road databases. This paper applies the Winner Takes All (WTA) mechanism, derived from other research fields, to combine the benefits of pixel and region classification. We fuse LiDAR data and multi-spectral high-resolution images to generate the set of features used by boosted classifiers to detect buildings, trees, bare land and grass. The main benefit of region based classification is that it removes the sensibility to noise of pixel based classifiers. The WTA approach is useful especially when pixel based approaches leave many pixels unclassified; typical cases are borders of building roofs or thin canopies, where LiDAR data are often noisy. Results in an urban environment using high-resolution LiDAR and multi-spectral data are presented comparing the performance of pixel, region and WTA approaches.